1 /*
   2  * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "memory/allocation.inline.hpp"
  27 #include "opto/block.hpp"
  28 #include "opto/c2compiler.hpp"
  29 #include "opto/callnode.hpp"
  30 #include "opto/cfgnode.hpp"
  31 #include "opto/machnode.hpp"
  32 #include "opto/runtime.hpp"
  33 #ifdef TARGET_ARCH_MODEL_x86_32
  34 # include "adfiles/ad_x86_32.hpp"
  35 #endif
  36 #ifdef TARGET_ARCH_MODEL_x86_64
  37 # include "adfiles/ad_x86_64.hpp"
  38 #endif
  39 #ifdef TARGET_ARCH_MODEL_sparc
  40 # include "adfiles/ad_sparc.hpp"
  41 #endif
  42 #ifdef TARGET_ARCH_MODEL_zero
  43 # include "adfiles/ad_zero.hpp"
  44 #endif
  45 #ifdef TARGET_ARCH_MODEL_arm
  46 # include "adfiles/ad_arm.hpp"
  47 #endif
  48 #ifdef TARGET_ARCH_MODEL_ppc
  49 # include "adfiles/ad_ppc.hpp"
  50 #endif
  51 
  52 // Optimization - Graph Style
  53 
  54 //------------------------------implicit_null_check----------------------------
  55 // Detect implicit-null-check opportunities.  Basically, find NULL checks
  56 // with suitable memory ops nearby.  Use the memory op to do the NULL check.
  57 // I can generate a memory op if there is not one nearby.
  58 // The proj is the control projection for the not-null case.
  59 // The val is the pointer being checked for nullness or
  60 // decodeHeapOop_not_null node if it did not fold into address.
  61 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
  62   // Assume if null check need for 0 offset then always needed
  63   // Intel solaris doesn't support any null checks yet and no
  64   // mechanism exists (yet) to set the switches at an os_cpu level
  65   if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
  66 
  67   // Make sure the ptr-is-null path appears to be uncommon!
  68   float f = end()->as_MachIf()->_prob;
  69   if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
  70   if( f > PROB_UNLIKELY_MAG(4) ) return;
  71 
  72   uint bidx = 0;                // Capture index of value into memop
  73   bool was_store;               // Memory op is a store op
  74 
  75   // Get the successor block for if the test ptr is non-null
  76   Block* not_null_block;  // this one goes with the proj
  77   Block* null_block;
  78   if (_nodes[_nodes.size()-1] == proj) {
  79     null_block     = _succs[0];
  80     not_null_block = _succs[1];
  81   } else {
  82     assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
  83     not_null_block = _succs[0];
  84     null_block     = _succs[1];
  85   }
  86   while (null_block->is_Empty() == Block::empty_with_goto) {
  87     null_block     = null_block->_succs[0];
  88   }
  89 
  90   // Search the exception block for an uncommon trap.
  91   // (See Parse::do_if and Parse::do_ifnull for the reason
  92   // we need an uncommon trap.  Briefly, we need a way to
  93   // detect failure of this optimization, as in 6366351.)
  94   {
  95     bool found_trap = false;
  96     for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
  97       Node* nn = null_block->_nodes[i1];
  98       if (nn->is_MachCall() &&
  99           nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
 100         const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
 101         if (trtype->isa_int() && trtype->is_int()->is_con()) {
 102           jint tr_con = trtype->is_int()->get_con();
 103           Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
 104           Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
 105           assert((int)reason < (int)BitsPerInt, "recode bit map");
 106           if (is_set_nth_bit(allowed_reasons, (int) reason)
 107               && action != Deoptimization::Action_none) {
 108             // This uncommon trap is sure to recompile, eventually.
 109             // When that happens, C->too_many_traps will prevent
 110             // this transformation from happening again.
 111             found_trap = true;
 112           }
 113         }
 114         break;
 115       }
 116     }
 117     if (!found_trap) {
 118       // We did not find an uncommon trap.
 119       return;
 120     }
 121   }
 122 
 123   // Check for decodeHeapOop_not_null node which did not fold into address
 124   bool is_decoden = ((intptr_t)val) & 1;
 125   val = (Node*)(((intptr_t)val) & ~1);
 126 
 127   assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
 128          (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
 129 
 130   // Search the successor block for a load or store who's base value is also
 131   // the tested value.  There may be several.
 132   Node_List *out = new Node_List(Thread::current()->resource_area());
 133   MachNode *best = NULL;        // Best found so far
 134   for (DUIterator i = val->outs(); val->has_out(i); i++) {
 135     Node *m = val->out(i);
 136     if( !m->is_Mach() ) continue;
 137     MachNode *mach = m->as_Mach();
 138     was_store = false;
 139     int iop = mach->ideal_Opcode();
 140     switch( iop ) {
 141     case Op_LoadB:
 142     case Op_LoadUB:
 143     case Op_LoadUS:
 144     case Op_LoadD:
 145     case Op_LoadF:
 146     case Op_LoadI:
 147     case Op_LoadL:
 148     case Op_LoadP:
 149     case Op_LoadN:
 150     case Op_LoadS:
 151     case Op_LoadKlass:
 152     case Op_LoadNKlass:
 153     case Op_LoadRange:
 154     case Op_LoadD_unaligned:
 155     case Op_LoadL_unaligned:
 156       assert(mach->in(2) == val, "should be address");
 157       break;
 158     case Op_StoreB:
 159     case Op_StoreC:
 160     case Op_StoreCM:
 161     case Op_StoreD:
 162     case Op_StoreF:
 163     case Op_StoreI:
 164     case Op_StoreL:
 165     case Op_StoreP:
 166     case Op_StoreN:
 167       was_store = true;         // Memory op is a store op
 168       // Stores will have their address in slot 2 (memory in slot 1).
 169       // If the value being nul-checked is in another slot, it means we
 170       // are storing the checked value, which does NOT check the value!
 171       if( mach->in(2) != val ) continue;
 172       break;                    // Found a memory op?
 173     case Op_StrComp:
 174     case Op_StrEquals:
 175     case Op_StrIndexOf:
 176     case Op_AryEq:
 177       // Not a legit memory op for implicit null check regardless of
 178       // embedded loads
 179       continue;
 180     default:                    // Also check for embedded loads
 181       if( !mach->needs_anti_dependence_check() )
 182         continue;               // Not an memory op; skip it
 183       if( must_clone[iop] ) {
 184         // Do not move nodes which produce flags because
 185         // RA will try to clone it to place near branch and
 186         // it will cause recompilation, see clone_node().
 187         continue;
 188       }
 189       {
 190         // Check that value is used in memory address in
 191         // instructions with embedded load (CmpP val1,(val2+off)).
 192         Node* base;
 193         Node* index;
 194         const MachOper* oper = mach->memory_inputs(base, index);
 195         if (oper == NULL || oper == (MachOper*)-1) {
 196           continue;             // Not an memory op; skip it
 197         }
 198         if (val == base ||
 199             val == index && val->bottom_type()->isa_narrowoop()) {
 200           break;                // Found it
 201         } else {
 202           continue;             // Skip it
 203         }
 204       }
 205       break;
 206     }
 207     // check if the offset is not too high for implicit exception
 208     {
 209       intptr_t offset = 0;
 210       const TypePtr *adr_type = NULL;  // Do not need this return value here
 211       const Node* base = mach->get_base_and_disp(offset, adr_type);
 212       if (base == NULL || base == NodeSentinel) {
 213         // Narrow oop address doesn't have base, only index
 214         if( val->bottom_type()->isa_narrowoop() &&
 215             MacroAssembler::needs_explicit_null_check(offset) )
 216           continue;             // Give up if offset is beyond page size
 217         // cannot reason about it; is probably not implicit null exception
 218       } else {
 219         const TypePtr* tptr;
 220         if (UseCompressedOops && (Universe::narrow_oop_shift() == 0)) {
 221           // 32-bits narrow oop can be the base of address expressions
 222           tptr = base->get_ptr_type();
 223         } else {
 224           // only regular oops are expected here
 225           tptr = base->bottom_type()->is_ptr();
 226         }
 227         // Give up if offset is not a compile-time constant
 228         if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
 229           continue;
 230         offset += tptr->_offset; // correct if base is offseted
 231         if( MacroAssembler::needs_explicit_null_check(offset) )
 232           continue;             // Give up is reference is beyond 4K page size
 233       }
 234     }
 235 
 236     // Check ctrl input to see if the null-check dominates the memory op
 237     Block *cb = cfg->_bbs[mach->_idx];
 238     cb = cb->_idom;             // Always hoist at least 1 block
 239     if( !was_store ) {          // Stores can be hoisted only one block
 240       while( cb->_dom_depth > (_dom_depth + 1))
 241         cb = cb->_idom;         // Hoist loads as far as we want
 242       // The non-null-block should dominate the memory op, too. Live
 243       // range spilling will insert a spill in the non-null-block if it is
 244       // needs to spill the memory op for an implicit null check.
 245       if (cb->_dom_depth == (_dom_depth + 1)) {
 246         if (cb != not_null_block) continue;
 247         cb = cb->_idom;
 248       }
 249     }
 250     if( cb != this ) continue;
 251 
 252     // Found a memory user; see if it can be hoisted to check-block
 253     uint vidx = 0;              // Capture index of value into memop
 254     uint j;
 255     for( j = mach->req()-1; j > 0; j-- ) {
 256       if( mach->in(j) == val ) {
 257         vidx = j;
 258         // Ignore DecodeN val which could be hoisted to where needed.
 259         if( is_decoden ) continue;
 260       }
 261       // Block of memory-op input
 262       Block *inb = cfg->_bbs[mach->in(j)->_idx];
 263       Block *b = this;          // Start from nul check
 264       while( b != inb && b->_dom_depth > inb->_dom_depth )
 265         b = b->_idom;           // search upwards for input
 266       // See if input dominates null check
 267       if( b != inb )
 268         break;
 269     }
 270     if( j > 0 )
 271       continue;
 272     Block *mb = cfg->_bbs[mach->_idx];
 273     // Hoisting stores requires more checks for the anti-dependence case.
 274     // Give up hoisting if we have to move the store past any load.
 275     if( was_store ) {
 276       Block *b = mb;            // Start searching here for a local load
 277       // mach use (faulting) trying to hoist
 278       // n might be blocker to hoisting
 279       while( b != this ) {
 280         uint k;
 281         for( k = 1; k < b->_nodes.size(); k++ ) {
 282           Node *n = b->_nodes[k];
 283           if( n->needs_anti_dependence_check() &&
 284               n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
 285             break;              // Found anti-dependent load
 286         }
 287         if( k < b->_nodes.size() )
 288           break;                // Found anti-dependent load
 289         // Make sure control does not do a merge (would have to check allpaths)
 290         if( b->num_preds() != 2 ) break;
 291         b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
 292       }
 293       if( b != this ) continue;
 294     }
 295 
 296     // Make sure this memory op is not already being used for a NullCheck
 297     Node *e = mb->end();
 298     if( e->is_MachNullCheck() && e->in(1) == mach )
 299       continue;                 // Already being used as a NULL check
 300 
 301     // Found a candidate!  Pick one with least dom depth - the highest
 302     // in the dom tree should be closest to the null check.
 303     if( !best ||
 304         cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
 305       best = mach;
 306       bidx = vidx;
 307 
 308     }
 309   }
 310   // No candidate!
 311   if( !best ) return;
 312 
 313   // ---- Found an implicit null check
 314   extern int implicit_null_checks;
 315   implicit_null_checks++;
 316 
 317   if( is_decoden ) {
 318     // Check if we need to hoist decodeHeapOop_not_null first.
 319     Block *valb = cfg->_bbs[val->_idx];
 320     if( this != valb && this->_dom_depth < valb->_dom_depth ) {
 321       // Hoist it up to the end of the test block.
 322       valb->find_remove(val);
 323       this->add_inst(val);
 324       cfg->_bbs.map(val->_idx,this);
 325       // DecodeN on x86 may kill flags. Check for flag-killing projections
 326       // that also need to be hoisted.
 327       for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
 328         Node* n = val->fast_out(j);
 329         if( n->is_MachProj() ) {
 330           cfg->_bbs[n->_idx]->find_remove(n);
 331           this->add_inst(n);
 332           cfg->_bbs.map(n->_idx,this);
 333         }
 334       }
 335     }
 336   }
 337   // Hoist the memory candidate up to the end of the test block.
 338   Block *old_block = cfg->_bbs[best->_idx];
 339   old_block->find_remove(best);
 340   add_inst(best);
 341   cfg->_bbs.map(best->_idx,this);
 342 
 343   // Move the control dependence
 344   if (best->in(0) && best->in(0) == old_block->_nodes[0])
 345     best->set_req(0, _nodes[0]);
 346 
 347   // Check for flag-killing projections that also need to be hoisted
 348   // Should be DU safe because no edge updates.
 349   for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
 350     Node* n = best->fast_out(j);
 351     if( n->is_MachProj() ) {
 352       cfg->_bbs[n->_idx]->find_remove(n);
 353       add_inst(n);
 354       cfg->_bbs.map(n->_idx,this);
 355     }
 356   }
 357 
 358   Compile *C = cfg->C;
 359   // proj==Op_True --> ne test; proj==Op_False --> eq test.
 360   // One of two graph shapes got matched:
 361   //   (IfTrue  (If (Bool NE (CmpP ptr NULL))))
 362   //   (IfFalse (If (Bool EQ (CmpP ptr NULL))))
 363   // NULL checks are always branch-if-eq.  If we see a IfTrue projection
 364   // then we are replacing a 'ne' test with a 'eq' NULL check test.
 365   // We need to flip the projections to keep the same semantics.
 366   if( proj->Opcode() == Op_IfTrue ) {
 367     // Swap order of projections in basic block to swap branch targets
 368     Node *tmp1 = _nodes[end_idx()+1];
 369     Node *tmp2 = _nodes[end_idx()+2];
 370     _nodes.map(end_idx()+1, tmp2);
 371     _nodes.map(end_idx()+2, tmp1);
 372     Node *tmp = new (C) Node(C->top()); // Use not NULL input
 373     tmp1->replace_by(tmp);
 374     tmp2->replace_by(tmp1);
 375     tmp->replace_by(tmp2);
 376     tmp->destruct();
 377   }
 378 
 379   // Remove the existing null check; use a new implicit null check instead.
 380   // Since schedule-local needs precise def-use info, we need to correct
 381   // it as well.
 382   Node *old_tst = proj->in(0);
 383   MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
 384   _nodes.map(end_idx(),nul_chk);
 385   cfg->_bbs.map(nul_chk->_idx,this);
 386   // Redirect users of old_test to nul_chk
 387   for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
 388     old_tst->last_out(i2)->set_req(0, nul_chk);
 389   // Clean-up any dead code
 390   for (uint i3 = 0; i3 < old_tst->req(); i3++)
 391     old_tst->set_req(i3, NULL);
 392 
 393   cfg->latency_from_uses(nul_chk);
 394   cfg->latency_from_uses(best);
 395 }
 396 
 397 
 398 //------------------------------select-----------------------------------------
 399 // Select a nice fellow from the worklist to schedule next. If there is only
 400 // one choice, then use it. Projections take top priority for correctness
 401 // reasons - if I see a projection, then it is next.  There are a number of
 402 // other special cases, for instructions that consume condition codes, et al.
 403 // These are chosen immediately. Some instructions are required to immediately
 404 // precede the last instruction in the block, and these are taken last. Of the
 405 // remaining cases (most), choose the instruction with the greatest latency
 406 // (that is, the most number of pseudo-cycles required to the end of the
 407 // routine). If there is a tie, choose the instruction with the most inputs.
 408 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
 409 
 410   // If only a single entry on the stack, use it
 411   uint cnt = worklist.size();
 412   if (cnt == 1) {
 413     Node *n = worklist[0];
 414     worklist.map(0,worklist.pop());
 415     return n;
 416   }
 417 
 418   uint choice  = 0; // Bigger is most important
 419   uint latency = 0; // Bigger is scheduled first
 420   uint score   = 0; // Bigger is better
 421   int idx = -1;     // Index in worklist
 422 
 423   for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
 424     // Order in worklist is used to break ties.
 425     // See caller for how this is used to delay scheduling
 426     // of induction variable increments to after the other
 427     // uses of the phi are scheduled.
 428     Node *n = worklist[i];      // Get Node on worklist
 429 
 430     int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
 431     if( n->is_Proj() ||         // Projections always win
 432         n->Opcode()== Op_Con || // So does constant 'Top'
 433         iop == Op_CreateEx ||   // Create-exception must start block
 434         iop == Op_CheckCastPP
 435         ) {
 436       worklist.map(i,worklist.pop());
 437       return n;
 438     }
 439 
 440     // Final call in a block must be adjacent to 'catch'
 441     Node *e = end();
 442     if( e->is_Catch() && e->in(0)->in(0) == n )
 443       continue;
 444 
 445     // Memory op for an implicit null check has to be at the end of the block
 446     if( e->is_MachNullCheck() && e->in(1) == n )
 447       continue;
 448 
 449     // Schedule IV increment last.
 450     if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
 451         e->in(1)->in(1) == n && n->is_iteratively_computed())
 452       continue;
 453 
 454     uint n_choice  = 2;
 455 
 456     // See if this instruction is consumed by a branch. If so, then (as the
 457     // branch is the last instruction in the basic block) force it to the
 458     // end of the basic block
 459     if ( must_clone[iop] ) {
 460       // See if any use is a branch
 461       bool found_machif = false;
 462 
 463       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 464         Node* use = n->fast_out(j);
 465 
 466         // The use is a conditional branch, make them adjacent
 467         if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
 468           found_machif = true;
 469           break;
 470         }
 471 
 472         // More than this instruction pending for successor to be ready,
 473         // don't choose this if other opportunities are ready
 474         if (ready_cnt.at(use->_idx) > 1)
 475           n_choice = 1;
 476       }
 477 
 478       // loop terminated, prefer not to use this instruction
 479       if (found_machif)
 480         continue;
 481     }
 482 
 483     // See if this has a predecessor that is "must_clone", i.e. sets the
 484     // condition code. If so, choose this first
 485     for (uint j = 0; j < n->req() ; j++) {
 486       Node *inn = n->in(j);
 487       if (inn) {
 488         if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
 489           n_choice = 3;
 490           break;
 491         }
 492       }
 493     }
 494 
 495     // MachTemps should be scheduled last so they are near their uses
 496     if (n->is_MachTemp()) {
 497       n_choice = 1;
 498     }
 499 
 500     uint n_latency = cfg->_node_latency->at_grow(n->_idx);
 501     uint n_score   = n->req();   // Many inputs get high score to break ties
 502 
 503     // Keep best latency found
 504     if( choice < n_choice ||
 505         ( choice == n_choice &&
 506           ( latency < n_latency ||
 507             ( latency == n_latency &&
 508               ( score < n_score ))))) {
 509       choice  = n_choice;
 510       latency = n_latency;
 511       score   = n_score;
 512       idx     = i;               // Also keep index in worklist
 513     }
 514   } // End of for all ready nodes in worklist
 515 
 516   assert(idx >= 0, "index should be set");
 517   Node *n = worklist[(uint)idx];      // Get the winner
 518 
 519   worklist.map((uint)idx, worklist.pop());     // Compress worklist
 520   return n;
 521 }
 522 
 523 
 524 //------------------------------set_next_call----------------------------------
 525 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
 526   if( next_call.test_set(n->_idx) ) return;
 527   for( uint i=0; i<n->len(); i++ ) {
 528     Node *m = n->in(i);
 529     if( !m ) continue;  // must see all nodes in block that precede call
 530     if( bbs[m->_idx] == this )
 531       set_next_call( m, next_call, bbs );
 532   }
 533 }
 534 
 535 //------------------------------needed_for_next_call---------------------------
 536 // Set the flag 'next_call' for each Node that is needed for the next call to
 537 // be scheduled.  This flag lets me bias scheduling so Nodes needed for the
 538 // next subroutine call get priority - basically it moves things NOT needed
 539 // for the next call till after the call.  This prevents me from trying to
 540 // carry lots of stuff live across a call.
 541 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
 542   // Find the next control-defining Node in this block
 543   Node* call = NULL;
 544   for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
 545     Node* m = this_call->fast_out(i);
 546     if( bbs[m->_idx] == this && // Local-block user
 547         m != this_call &&       // Not self-start node
 548         m->is_MachCall() )
 549       call = m;
 550       break;
 551   }
 552   if (call == NULL)  return;    // No next call (e.g., block end is near)
 553   // Set next-call for all inputs to this call
 554   set_next_call(call, next_call, bbs);
 555 }
 556 
 557 //------------------------------add_call_kills-------------------------------------
 558 void Block::add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
 559   // Fill in the kill mask for the call
 560   for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
 561     if( !regs.Member(r) ) {     // Not already defined by the call
 562       // Save-on-call register?
 563       if ((save_policy[r] == 'C') ||
 564           (save_policy[r] == 'A') ||
 565           ((save_policy[r] == 'E') && exclude_soe)) {
 566         proj->_rout.Insert(r);
 567       }
 568     }
 569   }
 570 }
 571 
 572 
 573 //------------------------------sched_call-------------------------------------
 574 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
 575   RegMask regs;
 576 
 577   // Schedule all the users of the call right now.  All the users are
 578   // projection Nodes, so they must be scheduled next to the call.
 579   // Collect all the defined registers.
 580   for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
 581     Node* n = mcall->fast_out(i);
 582     assert( n->is_MachProj(), "" );
 583     int n_cnt = ready_cnt.at(n->_idx)-1;
 584     ready_cnt.at_put(n->_idx, n_cnt);
 585     assert( n_cnt == 0, "" );
 586     // Schedule next to call
 587     _nodes.map(node_cnt++, n);
 588     // Collect defined registers
 589     regs.OR(n->out_RegMask());
 590     // Check for scheduling the next control-definer
 591     if( n->bottom_type() == Type::CONTROL )
 592       // Warm up next pile of heuristic bits
 593       needed_for_next_call(n, next_call, bbs);
 594 
 595     // Children of projections are now all ready
 596     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 597       Node* m = n->fast_out(j); // Get user
 598       if( bbs[m->_idx] != this ) continue;
 599       if( m->is_Phi() ) continue;
 600       int m_cnt = ready_cnt.at(m->_idx)-1;
 601       ready_cnt.at_put(m->_idx, m_cnt);
 602       if( m_cnt == 0 )
 603         worklist.push(m);
 604     }
 605 
 606   }
 607 
 608   // Act as if the call defines the Frame Pointer.
 609   // Certainly the FP is alive and well after the call.
 610   regs.Insert(matcher.c_frame_pointer());
 611 
 612   // Set all registers killed and not already defined by the call.
 613   uint r_cnt = mcall->tf()->range()->cnt();
 614   int op = mcall->ideal_Opcode();
 615   MachProjNode *proj = new (matcher.C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
 616   bbs.map(proj->_idx,this);
 617   _nodes.insert(node_cnt++, proj);
 618 
 619   // Select the right register save policy.
 620   const char * save_policy;
 621   switch (op) {
 622     case Op_CallRuntime:
 623     case Op_CallLeaf:
 624     case Op_CallLeafNoFP:
 625       // Calling C code so use C calling convention
 626       save_policy = matcher._c_reg_save_policy;
 627       break;
 628 
 629     case Op_CallStaticJava:
 630     case Op_CallDynamicJava:
 631       // Calling Java code so use Java calling convention
 632       save_policy = matcher._register_save_policy;
 633       break;
 634 
 635     default:
 636       ShouldNotReachHere();
 637   }
 638 
 639   // When using CallRuntime mark SOE registers as killed by the call
 640   // so values that could show up in the RegisterMap aren't live in a
 641   // callee saved register since the register wouldn't know where to
 642   // find them.  CallLeaf and CallLeafNoFP are ok because they can't
 643   // have debug info on them.  Strictly speaking this only needs to be
 644   // done for oops since idealreg2debugmask takes care of debug info
 645   // references but there no way to handle oops differently than other
 646   // pointers as far as the kill mask goes.
 647   bool exclude_soe = op == Op_CallRuntime;
 648 
 649   // If the call is a MethodHandle invoke, we need to exclude the
 650   // register which is used to save the SP value over MH invokes from
 651   // the mask.  Otherwise this register could be used for
 652   // deoptimization information.
 653   if (op == Op_CallStaticJava) {
 654     MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
 655     if (mcallstaticjava->_method_handle_invoke)
 656       proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
 657   }
 658 
 659   add_call_kills(proj, regs, save_policy, exclude_soe);
 660 
 661   return node_cnt;
 662 }
 663 
 664 
 665 //------------------------------schedule_local---------------------------------
 666 // Topological sort within a block.  Someday become a real scheduler.
 667 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, GrowableArray<int> &ready_cnt, VectorSet &next_call) {
 668   // Already "sorted" are the block start Node (as the first entry), and
 669   // the block-ending Node and any trailing control projections.  We leave
 670   // these alone.  PhiNodes and ParmNodes are made to follow the block start
 671   // Node.  Everything else gets topo-sorted.
 672 
 673 #ifndef PRODUCT
 674     if (cfg->trace_opto_pipelining()) {
 675       tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
 676       for (uint i = 0;i < _nodes.size();i++) {
 677         tty->print("# ");
 678         _nodes[i]->fast_dump();
 679       }
 680       tty->print_cr("#");
 681     }
 682 #endif
 683 
 684   // RootNode is already sorted
 685   if( _nodes.size() == 1 ) return true;
 686 
 687   // Move PhiNodes and ParmNodes from 1 to cnt up to the start
 688   uint node_cnt = end_idx();
 689   uint phi_cnt = 1;
 690   uint i;
 691   for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
 692     Node *n = _nodes[i];
 693     if( n->is_Phi() ||          // Found a PhiNode or ParmNode
 694         (n->is_Proj()  && n->in(0) == head()) ) {
 695       // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
 696       _nodes.map(i,_nodes[phi_cnt]);
 697       _nodes.map(phi_cnt++,n);  // swap Phi/Parm up front
 698     } else {                    // All others
 699       // Count block-local inputs to 'n'
 700       uint cnt = n->len();      // Input count
 701       uint local = 0;
 702       for( uint j=0; j<cnt; j++ ) {
 703         Node *m = n->in(j);
 704         if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
 705           local++;              // One more block-local input
 706       }
 707       ready_cnt.at_put(n->_idx, local); // Count em up
 708 
 709 #ifdef ASSERT
 710       if( UseConcMarkSweepGC || UseG1GC ) {
 711         if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
 712           // Check the precedence edges
 713           for (uint prec = n->req(); prec < n->len(); prec++) {
 714             Node* oop_store = n->in(prec);
 715             if (oop_store != NULL) {
 716               assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
 717             }
 718           }
 719         }
 720       }
 721 #endif
 722 
 723       // A few node types require changing a required edge to a precedence edge
 724       // before allocation.
 725       if( n->is_Mach() && n->req() > TypeFunc::Parms &&
 726           (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
 727            n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
 728         // MemBarAcquire could be created without Precedent edge.
 729         // del_req() replaces the specified edge with the last input edge
 730         // and then removes the last edge. If the specified edge > number of
 731         // edges the last edge will be moved outside of the input edges array
 732         // and the edge will be lost. This is why this code should be
 733         // executed only when Precedent (== TypeFunc::Parms) edge is present.
 734         Node *x = n->in(TypeFunc::Parms);
 735         n->del_req(TypeFunc::Parms);
 736         n->add_prec(x);
 737       }
 738     }
 739   }
 740   for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
 741     ready_cnt.at_put(_nodes[i2]->_idx, 0);
 742 
 743   // All the prescheduled guys do not hold back internal nodes
 744   uint i3;
 745   for(i3 = 0; i3<phi_cnt; i3++ ) {  // For all pre-scheduled
 746     Node *n = _nodes[i3];       // Get pre-scheduled
 747     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 748       Node* m = n->fast_out(j);
 749       if( cfg->_bbs[m->_idx] ==this ) { // Local-block user
 750         int m_cnt = ready_cnt.at(m->_idx)-1;
 751         ready_cnt.at_put(m->_idx, m_cnt);   // Fix ready count
 752       }
 753     }
 754   }
 755 
 756   Node_List delay;
 757   // Make a worklist
 758   Node_List worklist;
 759   for(uint i4=i3; i4<node_cnt; i4++ ) {    // Put ready guys on worklist
 760     Node *m = _nodes[i4];
 761     if( !ready_cnt.at(m->_idx) ) {   // Zero ready count?
 762       if (m->is_iteratively_computed()) {
 763         // Push induction variable increments last to allow other uses
 764         // of the phi to be scheduled first. The select() method breaks
 765         // ties in scheduling by worklist order.
 766         delay.push(m);
 767       } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
 768         // Force the CreateEx to the top of the list so it's processed
 769         // first and ends up at the start of the block.
 770         worklist.insert(0, m);
 771       } else {
 772         worklist.push(m);         // Then on to worklist!
 773       }
 774     }
 775   }
 776   while (delay.size()) {
 777     Node* d = delay.pop();
 778     worklist.push(d);
 779   }
 780 
 781   // Warm up the 'next_call' heuristic bits
 782   needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
 783 
 784 #ifndef PRODUCT
 785     if (cfg->trace_opto_pipelining()) {
 786       for (uint j=0; j<_nodes.size(); j++) {
 787         Node     *n = _nodes[j];
 788         int     idx = n->_idx;
 789         tty->print("#   ready cnt:%3d  ", ready_cnt.at(idx));
 790         tty->print("latency:%3d  ", cfg->_node_latency->at_grow(idx));
 791         tty->print("%4d: %s\n", idx, n->Name());
 792       }
 793     }
 794 #endif
 795 
 796   uint max_idx = (uint)ready_cnt.length();
 797   // Pull from worklist and schedule
 798   while( worklist.size() ) {    // Worklist is not ready
 799 
 800 #ifndef PRODUCT
 801     if (cfg->trace_opto_pipelining()) {
 802       tty->print("#   ready list:");
 803       for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
 804         Node *n = worklist[i];      // Get Node on worklist
 805         tty->print(" %d", n->_idx);
 806       }
 807       tty->cr();
 808     }
 809 #endif
 810 
 811     // Select and pop a ready guy from worklist
 812     Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
 813     _nodes.map(phi_cnt++,n);    // Schedule him next
 814 
 815 #ifndef PRODUCT
 816     if (cfg->trace_opto_pipelining()) {
 817       tty->print("#    select %d: %s", n->_idx, n->Name());
 818       tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx));
 819       n->dump();
 820       if (Verbose) {
 821         tty->print("#   ready list:");
 822         for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
 823           Node *n = worklist[i];      // Get Node on worklist
 824           tty->print(" %d", n->_idx);
 825         }
 826         tty->cr();
 827       }
 828     }
 829 
 830 #endif
 831     if( n->is_MachCall() ) {
 832       MachCallNode *mcall = n->as_MachCall();
 833       phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
 834       continue;
 835     }
 836 
 837     if (n->is_Mach() && n->as_Mach()->has_call()) {
 838       RegMask regs;
 839       regs.Insert(matcher.c_frame_pointer());
 840       regs.OR(n->out_RegMask());
 841 
 842       MachProjNode *proj = new (matcher.C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
 843       cfg->_bbs.map(proj->_idx,this);
 844       _nodes.insert(phi_cnt++, proj);
 845 
 846       add_call_kills(proj, regs, matcher._c_reg_save_policy, false);
 847     }
 848 
 849     // Children are now all ready
 850     for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
 851       Node* m = n->fast_out(i5); // Get user
 852       if( cfg->_bbs[m->_idx] != this ) continue;
 853       if( m->is_Phi() ) continue;
 854       if (m->_idx >= max_idx) { // new node, skip it
 855         assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
 856         continue;
 857       }
 858       int m_cnt = ready_cnt.at(m->_idx)-1;
 859       ready_cnt.at_put(m->_idx, m_cnt);
 860       if( m_cnt == 0 )
 861         worklist.push(m);
 862     }
 863   }
 864 
 865   if( phi_cnt != end_idx() ) {
 866     // did not schedule all.  Retry, Bailout, or Die
 867     Compile* C = matcher.C;
 868     if (C->subsume_loads() == true && !C->failing()) {
 869       // Retry with subsume_loads == false
 870       // If this is the first failure, the sentinel string will "stick"
 871       // to the Compile object, and the C2Compiler will see it and retry.
 872       C->record_failure(C2Compiler::retry_no_subsuming_loads());
 873     }
 874     // assert( phi_cnt == end_idx(), "did not schedule all" );
 875     return false;
 876   }
 877 
 878 #ifndef PRODUCT
 879   if (cfg->trace_opto_pipelining()) {
 880     tty->print_cr("#");
 881     tty->print_cr("# after schedule_local");
 882     for (uint i = 0;i < _nodes.size();i++) {
 883       tty->print("# ");
 884       _nodes[i]->fast_dump();
 885     }
 886     tty->cr();
 887   }
 888 #endif
 889 
 890 
 891   return true;
 892 }
 893 
 894 //--------------------------catch_cleanup_fix_all_inputs-----------------------
 895 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
 896   for (uint l = 0; l < use->len(); l++) {
 897     if (use->in(l) == old_def) {
 898       if (l < use->req()) {
 899         use->set_req(l, new_def);
 900       } else {
 901         use->rm_prec(l);
 902         use->add_prec(new_def);
 903         l--;
 904       }
 905     }
 906   }
 907 }
 908 
 909 //------------------------------catch_cleanup_find_cloned_def------------------
 910 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
 911   assert( use_blk != def_blk, "Inter-block cleanup only");
 912 
 913   // The use is some block below the Catch.  Find and return the clone of the def
 914   // that dominates the use. If there is no clone in a dominating block, then
 915   // create a phi for the def in a dominating block.
 916 
 917   // Find which successor block dominates this use.  The successor
 918   // blocks must all be single-entry (from the Catch only; I will have
 919   // split blocks to make this so), hence they all dominate.
 920   while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
 921     use_blk = use_blk->_idom;
 922 
 923   // Find the successor
 924   Node *fixup = NULL;
 925 
 926   uint j;
 927   for( j = 0; j < def_blk->_num_succs; j++ )
 928     if( use_blk == def_blk->_succs[j] )
 929       break;
 930 
 931   if( j == def_blk->_num_succs ) {
 932     // Block at same level in dom-tree is not a successor.  It needs a
 933     // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
 934     Node_Array inputs = new Node_List(Thread::current()->resource_area());
 935     for(uint k = 1; k < use_blk->num_preds(); k++) {
 936       inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
 937     }
 938 
 939     // Check to see if the use_blk already has an identical phi inserted.
 940     // If it exists, it will be at the first position since all uses of a
 941     // def are processed together.
 942     Node *phi = use_blk->_nodes[1];
 943     if( phi->is_Phi() ) {
 944       fixup = phi;
 945       for (uint k = 1; k < use_blk->num_preds(); k++) {
 946         if (phi->in(k) != inputs[k]) {
 947           // Not a match
 948           fixup = NULL;
 949           break;
 950         }
 951       }
 952     }
 953 
 954     // If an existing PhiNode was not found, make a new one.
 955     if (fixup == NULL) {
 956       Node *new_phi = PhiNode::make(use_blk->head(), def);
 957       use_blk->_nodes.insert(1, new_phi);
 958       bbs.map(new_phi->_idx, use_blk);
 959       for (uint k = 1; k < use_blk->num_preds(); k++) {
 960         new_phi->set_req(k, inputs[k]);
 961       }
 962       fixup = new_phi;
 963     }
 964 
 965   } else {
 966     // Found the use just below the Catch.  Make it use the clone.
 967     fixup = use_blk->_nodes[n_clone_idx];
 968   }
 969 
 970   return fixup;
 971 }
 972 
 973 //--------------------------catch_cleanup_intra_block--------------------------
 974 // Fix all input edges in use that reference "def".  The use is in the same
 975 // block as the def and both have been cloned in each successor block.
 976 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
 977 
 978   // Both the use and def have been cloned. For each successor block,
 979   // get the clone of the use, and make its input the clone of the def
 980   // found in that block.
 981 
 982   uint use_idx = blk->find_node(use);
 983   uint offset_idx = use_idx - beg;
 984   for( uint k = 0; k < blk->_num_succs; k++ ) {
 985     // Get clone in each successor block
 986     Block *sb = blk->_succs[k];
 987     Node *clone = sb->_nodes[offset_idx+1];
 988     assert( clone->Opcode() == use->Opcode(), "" );
 989 
 990     // Make use-clone reference the def-clone
 991     catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
 992   }
 993 }
 994 
 995 //------------------------------catch_cleanup_inter_block---------------------
 996 // Fix all input edges in use that reference "def".  The use is in a different
 997 // block than the def.
 998 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
 999   if( !use_blk ) return;        // Can happen if the use is a precedence edge
1000 
1001   Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
1002   catch_cleanup_fix_all_inputs(use, def, new_def);
1003 }
1004 
1005 //------------------------------call_catch_cleanup-----------------------------
1006 // If we inserted any instructions between a Call and his CatchNode,
1007 // clone the instructions on all paths below the Catch.
1008 void Block::call_catch_cleanup(Block_Array &bbs, Compile* C) {
1009 
1010   // End of region to clone
1011   uint end = end_idx();
1012   if( !_nodes[end]->is_Catch() ) return;
1013   // Start of region to clone
1014   uint beg = end;
1015   while(!_nodes[beg-1]->is_MachProj() ||
1016         !_nodes[beg-1]->in(0)->is_MachCall() ) {
1017     beg--;
1018     assert(beg > 0,"Catch cleanup walking beyond block boundary");
1019   }
1020   // Range of inserted instructions is [beg, end)
1021   if( beg == end ) return;
1022 
1023   // Clone along all Catch output paths.  Clone area between the 'beg' and
1024   // 'end' indices.
1025   for( uint i = 0; i < _num_succs; i++ ) {
1026     Block *sb = _succs[i];
1027     // Clone the entire area; ignoring the edge fixup for now.
1028     for( uint j = end; j > beg; j-- ) {
1029       // It is safe here to clone a node with anti_dependence
1030       // since clones dominate on each path.
1031       Node *clone = _nodes[j-1]->clone();
1032       sb->_nodes.insert( 1, clone );
1033       bbs.map(clone->_idx,sb);
1034     }
1035   }
1036 
1037 
1038   // Fixup edges.  Check the def-use info per cloned Node
1039   for(uint i2 = beg; i2 < end; i2++ ) {
1040     uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1041     Node *n = _nodes[i2];        // Node that got cloned
1042     // Need DU safe iterator because of edge manipulation in calls.
1043     Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1044     for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1045       out->push(n->fast_out(j1));
1046     }
1047     uint max = out->size();
1048     for (uint j = 0; j < max; j++) {// For all users
1049       Node *use = out->pop();
1050       Block *buse = bbs[use->_idx];
1051       if( use->is_Phi() ) {
1052         for( uint k = 1; k < use->req(); k++ )
1053           if( use->in(k) == n ) {
1054             Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
1055             use->set_req(k, fixup);
1056           }
1057       } else {
1058         if (this == buse) {
1059           catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
1060         } else {
1061           catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
1062         }
1063       }
1064     } // End for all users
1065 
1066   } // End of for all Nodes in cloned area
1067 
1068   // Remove the now-dead cloned ops
1069   for(uint i3 = beg; i3 < end; i3++ ) {
1070     _nodes[beg]->disconnect_inputs(NULL, C);
1071     _nodes.remove(beg);
1072   }
1073 
1074   // If the successor blocks have a CreateEx node, move it back to the top
1075   for(uint i4 = 0; i4 < _num_succs; i4++ ) {
1076     Block *sb = _succs[i4];
1077     uint new_cnt = end - beg;
1078     // Remove any newly created, but dead, nodes.
1079     for( uint j = new_cnt; j > 0; j-- ) {
1080       Node *n = sb->_nodes[j];
1081       if (n->outcnt() == 0 &&
1082           (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1083         n->disconnect_inputs(NULL, C);
1084         sb->_nodes.remove(j);
1085         new_cnt--;
1086       }
1087     }
1088     // If any newly created nodes remain, move the CreateEx node to the top
1089     if (new_cnt > 0) {
1090       Node *cex = sb->_nodes[1+new_cnt];
1091       if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1092         sb->_nodes.remove(1+new_cnt);
1093         sb->_nodes.insert(1,cex);
1094       }
1095     }
1096   }
1097 }